High-resolution crop map generation in Mediterranean environments using IOTA2 chain

Reliable crop mapping is essential in land and water management studies to understand the spatial distribution and dynamics of agricultural practices, to model resource use and production, and to propose sustainable scenarios for agricultural and water management. This work presents high-resolution crop mapping for the Mediterranean area, which is particularly interesting due to the limited data availability and the high level of land use heterogeneity. The main European land use dataset, Corine Land Cover (CLC), lacks the specificity required for accurate agricultural classification, especially for crop differentiation, and does not provide frequent or timely updates, which are crucial for many applications. Other more recent EU-wide crop mapping efforts (d’Andrimont et al, 2021) still lack regional accuracy due to widely scattered training data. To overcome these limitations, a large-scale crop mapping initiative was implemented in Sardinia to test and validate an artificial intelligence–based approach for Mediterranean environments. In this context, irrigated agriculture is a key sector for the sustainable management of limited water resources. The method uses Sentinel‑2 time series and survey data from the LPIS (Land Parcel Identification System). The study relies on IOTA², a land‑use map production chain first developed and tested at the French level, producing maps with 24 land‑use classes. The originality of the approach lies in the use of open‑source satellite images and an automated processing workflow based on supervised classifiers, making crop mapping faster and easily reproducible across years.

Learning samples are derived from 2018 LPIS data, supplemented by CLC and CLCplus Backbone datasets for natural areas and the Urban Atlas for urban areas. Two nomenclatures are tested: a detailed versus a simplified one, with 32 and 26 thematic classes, respectively, both focusing on Mediterranean-relevant crop typologies. The two nomenclatures are evaluated with sampling rates of 10%, 50%, and 100% of training pixels. Results show that the simplified nomenclature achieves higher accuracy, with an Overall Accuracy (OA) of 0.77 compared to 0.61 for the detailed nomenclature, using 100% training pixels. Increasing the training sample rate improves classification quality in both nomenclatures: in the short nomenclature, OA values are 0.596, 0.613, and 0.774 for 10%, 50%, and 100% sampling rates. In the detailed nomenclature, the improvement is weaker, with OA values of 0.596, 0.601, and 0.610, indicating that increasing sample size does not resolve class confusion. Among agricultural classes, rice, citrus, vegetables, and grapevine achieve the highest classification scores, which are among the crops with the largest irrigation requirements. Nuts, cereals, and fruit trees perform poorly, mainly due to insufficient training samples. Overall, the proposed nomenclature significantly improves the crop classes available in the CLC by increasing crop specificity and differentiation. This study presents a framework for fully automatic crop‑map production in Mediterranean environments, ensuring fast reproducibility over the years thanks to the use of openly accessible satellite imagery and an automated processing chain. This can improve the accuracy and reliability of water accounting for the agricultural sector and help promote sustainable use of limited water resources in the Mediterranean areas.